disambiguating the ambiguity disadvantage effect...
TRANSCRIPT
This is a repository copy of Disambiguating the ambiguity disadvantage effect: Behavioral and electrophysiological evidence for semantic competition.
White Rose Research Online URL for this paper:http://eprints.whiterose.ac.uk/158320/
Version: Accepted Version
Article:
Maciejewski, G and Klepousniotou, E orcid.org/0000-0002-2318-0951 (2020) Disambiguating the ambiguity disadvantage effect: Behavioral and electrophysiological evidence for semantic competition. Journal of Experimental Psychology: Learning, Memory, and Cognition. ISSN 0278-7393
https://doi.org/10.1037/xlm0000842
© 2020 American Psychological Association. This paper is not the copy of record and may not exactly replicate the authoritative document published in the APA journal. Please do not copy or cite without author's permission. The final article is available, upon publication, at: https://psycnet.apa.org/doi/10.1037/xlm0000842
[email protected]://eprints.whiterose.ac.uk/
Reuse
Items deposited in White Rose Research Online are protected by copyright, with all rights reserved unless indicated otherwise. They may be downloaded and/or printed for private study, or other acts as permitted by national copyright laws. The publisher or other rights holders may allow further reproduction and re-use of the full text version. This is indicated by the licence information on the White Rose Research Online record for the item.
Takedown
If you consider content in White Rose Research Online to be in breach of UK law, please notify us by emailing [email protected] including the URL of the record and the reason for the withdrawal request.
AMBIGUITY DISADVANTAGE 1
Running head: AMBIGUITY DISADVANTAGE
Disambiguating the ambiguity disadvantage effect:
Behavioral and electrophysiological evidence for
semantic competition
Greg Maciejewski1,2 & Ekaterini Klepousniotou1
1 School of Psychology, University of Leeds, UK
2 School of Education and Social Sciences, University of the West of Scotland, UK
Corresponding author:
Dr. Ekaterini Klepousniotou
School of Psychology
University of Leeds
Leeds
LS2 9JT
Phone: +44 (0)113 3435716
Email: [email protected]
AMBIGUITY DISADVANTAGE 2
Abstract
Semantic ambiguity has been shown to slow comprehension, though it is
unclear whether this “ambiguity disadvantage” is due to competition in semantic
activation or difficulties in response selection. We tested the two accounts by
examining semantic relatedness decisions to homonyms, or words with multiple
unrelated meanings (e.g., “football/electric fan”). Our behavioral results showed that
the ambiguity disadvantage arises only when the different meanings of words are of
comparable frequency, and are thus activated in parallel. Critically, this effect was
observed regardless of response-selection difficulties, both when the different
meanings triggered inconsistent responses on related trials (e.g., “fan-breeze”) and
consistent responses on unrelated trials (e.g., “fan-snake”). Our electrophysiological
results confirmed that this effect arises during semantic activation of the ambiguous
word, indexed by the N400, not during response selection. Overall, the findings show
that ambiguity resolution involves semantic competition and delineate why and when
this competition arises.
Keywords: lexical/semantic ambiguity; homonymy; meaning frequency; semantic processing; N400
AMBIGUITY DISADVANTAGE 3
1 Introduction
The vast majority of the words we use are in some way ambiguous, hence the
ability to select a single, contextually appropriate meaning without being overtly
distracted by the myriad of other possible meanings is a crucial component of any
theory of language comprehension. Indeed, the importance of understanding how
multiple meanings are represented and accessed is highlighted by the extensive
literature dedicated to this issue over the past few decades (for a recent review, see
Rodd, 2018).
One unclear finding in this literature is that of slower response/reading times
for ambiguous versus unambiguous words in tasks that require meaning selection in
neutral context or isolation. This so-called “ambiguity disadvantage” effect has been
typically observed for homonyms, words with multiple unrelated meanings, either in
late-disambiguation sentence reading (e.g., “He found the coach was too hot to
sleep in”; Duffy, Morris, & Rayner, 1988; Frazier & Rayner, 1990; Rayner & Duffy,
1986) or semantic relatedness decisions (e.g., “hide-conceal/skin”; Gottlob,
Goldinger, Stone, & Van Orden, 1999; Hoffman & Woollams, 2015; Pexman, Hino, &
Lupker, 2004; Piercey & Joordens, 2000). Although the effect appears to be robust,
little is still known as to why it arises, and what it reveals about the representations
and processes involved in ambiguity resolution. Here, we focus on homonyms and
examine two prominent accounts of the effect – semantic competition and decision
making.
The ambiguity disadvantage is an inherent prediction of the “distributed” view
of lexical-semantic representation (for an overview, see Seidenberg, 2007). In short,
connectionist models postulate that words are represented by units corresponding to
AMBIGUITY DISADVANTAGE 4
their orthographic and semantic features. These units are distributed, in the sense
that a single unit contributes to the representation of multiple words that share the
same feature. There are connections among the orthographic and semantic units
which, as a result of learning, acquire different weights reflecting the appropriate
form-to-meaning mapping. Thus, within this framework, it is the weights on the
connections that determine the ease of semantic activation.
For unambiguous words, the orthographic pattern of activation is always
associated with the same semantic pattern, which strengthens the connections and
facilitates future form-to-meaning mapping. Ambiguity, on the contrary, precludes
such a benefit. The orthographic pattern for words such as “bank” is ambiguous and
gives rise to a “blend state”, or partial activation of the different semantic
representations (“money/river bank”). As semantic activation increases, the
representations begin to compete for full activation, as only one of them can be
activated to complete the disambiguation process. According to connectionist
models of ambiguity (e.g., Armstrong & Plaut, 2008; Kawamoto, 1993; Rodd,
Gaskell, & Marslen-Wilson, 2004), it is this semantic competition, due to multiple
form-to-meaning mappings, that may account for the ambiguity disadvantage in word
comprehension.
The semantic competition account proposed in the connectionist models has
been challenged by Pexman et al. (2004), who argued that the ambiguity
disadvantage is due to decision-making difficulties in response selection. Their
semantic relatedness decision tasks revealed a substantial processing cost for
ambiguous words on related trials, regardless of which meaning the targets
instantiated (e.g., “hide-conceal/skin”). Interestingly, there was no such cost on
AMBIGUITY DISADVANTAGE 5
unrelated trials, where the same words were paired with unrelated targets (e.g.,
“hide-glass”). Pexman et al. (2004) reasoned that if the ambiguity disadvantage were
due to semantic activation processes, its effects would be observed on both related
and unrelated trials. On related trials, participants need to resolve ambiguity because
a blend state is not sufficient to support a correct response (e.g., only one of the
meanings of “hide” is related to “conceal”). On unrelated trials, participants may not
need to resolve ambiguity (e.g., both meanings of “hide” are unrelated to “glass”), but
their response is still entirely based on semantic activation. To accommodate their
findings, Pexman et al. (2004) posited that the processing cost specific to related
trials may be a task artefact caused by response conflict. Since the different
meanings of homonyms are inconsistent with the same response to a related target,
the cost may reflect the need to decide on which meaning should serve as response
input. Critically, no such response conflict arises on unrelated trials, hence the null
ambiguity effect when making relatedness decisions to unrelated word pairs.
Further support for the idea that the ambiguity disadvantage lies in decision-
making difficulties comes from Hino, Pexman, and Lupker’s (2006) semantic
categorisation studies. Since the different meanings of ambiguous words often fall
into different categories (e.g., “crane” in reference to the living/non-living category),
and may therefore create response conflict similar to that in relatedness decision
tasks, the researchers focused on “no” trials where neither meaning fell into a
category in question (e.g., “bear” in reference to the vegetable category). Their
results showed a processing cost for homonyms when the task involved broad living-
object or human-related categories, but not when it involved narrow animal or
vegetable categories. Hino et al. (2006) attributed this pattern of responses to the
AMBIGUITY DISADVANTAGE 6
nature of the decision category (see also Hargreaves, Pexman, Pittman, &
Goodyear, 2011). When the category is broad, participants must retrieve a large
number of semantic features of the target word’s referent and decide whether any of
them is true of the category. For ambiguous words, this may take considerably
longer because participants need to retrieve and analyze features of multiple word
referents, in case one of them falls into the category in question. In contrast, when
the category is well-defined and narrow, participants may be able to respond based
on a small number of features that are likely true of all the word referents, whilst
ignoring irrelevant features that would otherwise slow processing. The overall
argument, then, is that the ambiguity disadvantage arises only when task-relevant
decisions are somewhat more difficult to make.
In summary, the challenge of explaining the ambiguity disadvantage has
provided a strong impetus to the development of different accounts of ambiguity
representation. Under the semantic competition account (Armstrong & Plaut, 2008;
Kawamoto, 1993; Rodd et al., 2004), the delay in comprehension arises because
ambiguous words, in particular homonyms, have separate semantic representations
that compete for activation. Under the decision-making account (Hino et al., 2006;
Pexman et al., 2004), the delay arises due to task-specific response-selection
demands. The semantic representations are assumed to be activated independently,
without giving rise to any interference. Pexman et al. (2004) argued that such an
explanation would hold true if we assumed that the different meanings of ambiguous
words are represented in separate subsets of semantic memory, such that, for
example, the institution-related meaning of the word “bank” is represented within one
semantic space, whereas the river-related meaning is represented within another.
AMBIGUITY DISADVANTAGE 7
Overall, then, there is a clear need to establish the locus of the ambiguity
disadvantage before we make any further inferences about the structure of the
mental lexicon.
The present study directly tested the semantic competition and decision-
making accounts by investigating the ambiguity disadvantage in semantic
relatedness decision tasks, in which ambiguous/unambiguous words were followed
by related/unrelated targets. For ambiguous words, we focused on homonyms,
expecting that if any form of ambiguity produced competition at the semantic level, it
would be, foremost, observed for homonyms whose unrelated meanings are
unanimously assumed to have separate semantic representations (for a review, see
Eddington & Tokowicz, 2015). In Experiment 1, we contrasted homonymous and
non-homonymous words on related and unrelated trials to replicate the ambiguity
disadvantage in the first place. In Experiment 2, we used EEG measurements to
determine when this effect is in play – in other words, whether it arises during the
processing of the ambiguous word itself, as predicted by the semantic competition
account, or during response selection upon the presentation of the target, as
suggested by the decision-making account.
2 Experiment 1
Unlike previous relatedness decision studies (Gottlob et al., 1990; Pexman et
al., 2004; Piercey & Joordens, 2000), Experiment 1 made a clear distinction between
homonyms with balanced (e.g., “football/electric fan”) and unbalanced meaning
frequencies (e.g., “blue/spacious pen”). The rationale was that although all meanings
AMBIGUITY DISADVANTAGE 8
seem to be activated upon reading an ambiguous word, the broader literature on
ambiguity resolution indicates that the level and time-course of this activation are
influenced by meaning frequency, or dominance (for a review, see Twilley & Dixon,
2000). In particular, when ambiguous words are encountered in isolation or neutral
context (as in the present study), readers are biased towards the high-frequency
(HF) meaning, such that activation of the low-frequency (LF) counterpart is
noticeably weaker, delayed, and transient (Frost & Bentin, 1992; Klepousniotou,
Pike, Steinhauer, & Gracco, 2012; Simpson & Burgess, 1985). Drawing on this line
of research, we argue that any adequate account of how activation of multiple
meanings affects word comprehension must take into account the role of meaning
frequency in the activation process, or the distinction between balanced and
unbalanced homonyms. For example, if semantic competition does arise, one would
expect it to be maximal for balanced homonyms whose different meanings are
initially activated to the same extent and in parallel (in neutral or out of context). For
unbalanced homonyms, on the other hand, the impact of meaning frequency should
eliminate, or at least reduce, the competition. Readers may fully retrieve and select
the HF meaning very fast, such that the LF counterpart does not reach a sufficient
level of activation to engage in the competition.
The idea that meaning frequency modulates ambiguity effects in word
processing is not entirely new, as there have been a few studies that either
controlled for (Armstrong & Plaut, 2016; Mirman, Strauss, Dixon, & Magnuson, 2010)
or manipulated this property (Brocher, Koenig, Mauner, & Foraker, 2018; Grindrod,
Garnett, Malyutina, & den Ouden, 2014; Klepousniotou & Baum, 2007;
Klepousniotou et al., 2012; MacGregor, Bouwsema, & Klepousniotou, 2015). In
AMBIGUITY DISADVANTAGE 9
particular, Armstrong, Tokowicz, and Plaut (2012) suggested that the impact of
homonymy in word recognition depends on the relative frequencies of the multiple
meanings, such that there is a slight slowing in lexical decisions to balanced but not
unbalanced homonyms (but cf. Grindrod et al., 2014; Klepousniotou & Baum, 2007).
As for the impact in word comprehension, late-disambiguation sentence-reading
studies (Duffy et al., 1988; Rayner & Duffy, 1986) reported a similar pattern of results
- a processing cost for balanced but not unbalanced homonyms. This is in line with
Kawamoto’s (1993) model simulations predicting the ambiguity disadvantage to be
more pronounced when the different meanings are of comparable frequency, and
thus equal competitors in the race for activation.
Taken together, we sought to replicate and further examine the ambiguity
disadvantage for balanced but not unbalanced homonyms in semantic relatedness
decisions. This was necessary for Experiment 2 where we used EEG measurements
to establish when and why the disadvantage arises, separating early semantic
activation processes during prime presentation from late response-selection
processes during target presentation. Note that Experiment 1 involved two versions
that differed in the duration of the ambiguous/unambiguous prime (200 ms in
Experiment 1a, 700 ms in Experiment 1b). The rationale was that although studies
indicated a delay in LF-meaning activation for homonyms on the whole (Frost &
Bentin, 1992; Simpson & Burgess, 1985; Simpson & Krueger, 1891), it is unclear
how substantial this delay might be for highly unbalanced homonyms, such as those
used in the present study. Extending the prime duration was therefore essential to
confirm that unbalanced homonymy does not produce the ambiguity disadvantage,
AMBIGUITY DISADVANTAGE 10
either due to semantic competition or decision making, even when there is enough
time to retrieve the LF meaning.
2.1 Method
2.1.1 Participants
Participants were University of Leeds students and staff who participated for
course credit or £3. There were 35 participants [30 females, aged 19-35 (M = 25.8,
SD = 4.9)] in Experiment 1a and a different group of 30 [21 females, aged 18-42 (M
= 21.3, SD = 5.5)] in Experiment 1b. All participants were monolingual native
speakers of British English with no known history of language-/vision-related
difficulties or disorders. They were right-handed, as confirmed with the Briggs-Nebes
(1975) modified version of Annett’s (1967) handedness inventory. The experiment
received ethical approval from the School of Psychology, University of Leeds Ethics
Committee.
2.1.2 Stimuli
Prime words were 28 balanced homonyms (e.g., “fan), 28 unbalanced
homonyms (e.g., “pen”), and 56 non-homonyms (e.g., “crew”) that were split into two
sets (1 & 2). All homonyms were selected from the British norms of meaning
frequency (Maciejewski & Klepousniotou, 2016). The four sets of primes were
statistically comparable (all Fs < 1) with respect to 14 lexical and semantic variables,
such as form frequency and semantic diversity. For more information on prime-word
selection and matching, see Section 1.1 in the Supplementary Material.
AMBIGUITY DISADVANTAGE 11
We paired each homonymous prime with four target words: two semantically
related and two semantically unrelated. For unbalanced homonyms, one target
related to the HF meaning of the prime (e.g., “pen-ink”), while the other related to the
LF meaning (e.g., “pen-farmer”). The same manipulation was used for balanced
homonyms, although the difference in meaning frequencies for these items was
much smaller, as evident in the norms (Maciejewski & Klepousniotou, 2016). For
non-homonyms, the two related targets (A & B) referred to the same interpretation of
the prime (e.g., “fake-truth/fraud”). We also paired all primes with two targets (A & B)
that were unrelated to either of their meanings (e.g., “fan-snake/cancel”). This aimed
to equalize the number of related and unrelated word pairs in the experiment (all
listed in the Appendix). The 16 sets of targets were statistically comparable (all Fs <
1) with respect to 14 lexical and semantic variables, such as form frequency and
semantic diversity. For more information on target-word selection, matching, and
prime-target relatedness pre-test, see Sections 1.2 and 1.3 in the Supplementary
Material. For examples of different prime-target word pairs, see Table 1 below.
>> Insert Table 1 here <<
2.1.3 Procedure
The relatedness decision task was programmed in EPrime 2.0 (Schneider,
Eschman, & Zuccolotto, 2010). The task was to decide whether the prime and the
target were related in meaning by pressing a keyboard button. Participants pressed
the L button for “related” with their dominant (right) hand or the A button with their left
hand. Both response speed and accuracy were equally emphasized in the
AMBIGUITY DISADVANTAGE 12
instructions, and participants were instructed and given examples as to what
constitutes semantic relatedness.
The stimuli were pseudo-randomly divided into four blocks, such that each
block contained the same number of trials in the different conditions. Participants
responded to each prime four times, but none of the primes appeared more than
once within the same block. The order of blocks was counter-balanced across
participants. The order of trials within each block was pseudo-randomized, such that
no more than three related/unrelated word pairs appeared consecutively. The task
began with a practice block comprising two examples of each condition (N = 32) and
feedback on each response. There were two one-minute breaks - one after the
practice block and the other after the first two experimental blocks. Following each
break, participants first responded to eight filler trials (not included in the analysis)
that aimed to help them get back to the habit of quick responding.
In Experiment 1a, trials began with a 500 ms fixation cross. After 100 ms, the
prime and the target were presented for 200 ms and 500 ms, respectively, with a 50
ms interval in between. Once the target disappeared, there was 1500 ms for
response execution followed by a 100 ms inter-trial interval. Participants could make
relatedness decisions as soon as the target appeared, but they had to respond
within the first 1500 ms (i.e., responses of 1500-2000 ms were deemed too slow and
would be excluded from analyses). In Experiment 1b, the only difference was the
longer prime-word duration (700 ms instead of 200 ms).
2.2 Results
AMBIGUITY DISADVANTAGE 13
In Experiment 1a, two of the 35 participants were removed from analyses –
one due to a large number of errors on related trials (63.8%) and the other due to
slow and variable responses (M = 899.9, SD = 182.9). In Experiment 1b, one of the
30 participants was removed due to a large number of errors on related trials
(54.5%). In Experiment 1a, we also removed the four targets of one of the non-
homonyms as these were inadvertently paired with a different prime. For RTs, we
excluded errors (19% and 17% of trials in Experiments 1a and 1b, respectively) and
any responses that were two SDs above/below a participant’s mean in a given
condition (4% and 3% of trials in Experiments 1a and 1b, respectively). The
remaining RTs were log-transformed to further minimize the impact of potential
outliers and to normalize the residual distribution1.
Accuracy and latency data were analyzed using logit/linear mixed-effects
models with the factors of Prime (balanced homonym, unbalanced homonym, non-
homonym1, non-homonym2) and Target (HF-meaning/A, LF-meaning/B). RT models
also included Block (1, 2, 3, 4), though effects involving this factor are not reported
because its sole purpose was to account for potential practice or prime-repetition
effects, and no such effects were detected2 (Pollatsek & Well, 1995). Terms
involving Block were removed from accuracy models due to non-convergence.
Related and unrelated trials were analyzed separately, as our preliminary analyses
showed a significant effect of Trial (i.e., slower but more accurate responses on
unrelated trials) that always interacted with the effects of Prime and Target. These
preliminary analyses were otherwise the same as those reported below.
Furthermore, we were concerned that contrasts between ambiguity effects on related
and unrelated trials would not be readily interpretable, as it is not entirely clear what
AMBIGUITY DISADVANTAGE 14
processes underlie performance on unrelated trials, and whether/how they differ
from those on related trials.
Each model included significant random intercepts for subjects and items.
Following Barr, Levy, Scheepers, and Tily (2013) and Matuschek, Kliegl, Vasishth,
Baayen, and Bates (2017), the optimal random-effects structure justified by the data
was identified using forward model selection3. The only random slope that
significantly improved fit and was included in RT models was that of Block across
subjects. Fixed effects were tested using likelihood-ratio tests comparing full and
reduced models. All modelling was conducted using the “lme4” package (Bates,
Mächler, & Bolker, 2011) in R (R Development Core Team, 2004). Planned contrasts
examining the effects of Prime compared balanced/unbalanced homonyms to both
sets of non-homonyms. These tests were conducted using the “phia” package (De
Rosario-Martinez, 2015), and their significance threshold was adjusted using the
Holm-Bonferroni method to further prevent spurious results.
2.2.1 Related trials
In Experiment 1a, there was a significant main effect of Prime in both error
rates [ぬ2(3) = 63.1, p < .001] and RTs [ぬ2(3) = 39.1, p < .001]. Planned contrasts
showed less accurate and slower responses to both balanced (both ps < .001) and
unbalanced homonyms (both ps < .001) than to non-homonyms (see Figure 1
below). Responses were generally less accurate and slower to LF-meaning than HF-
meaning targets, and this main effect of Target [errors: ぬ2(1) = 37.9, p < .001; RTs:
ぬ2(1) = 13.7, p < .001] interacted with Prime [errors: ぬ2(3) = 39.7, p < .001; RTs: ぬ2(3)
= 26.2, p < .001]. Relative to both targets of non-homonyms, responses were less
AMBIGUITY DISADVANTAGE 15
accurate and slower to the LF-meaning targets of balanced (both ps < .001) and
unbalanced homonyms (both ps < .001) and the HF-meaning targets of balanced
homonyms (errors: p < .01; RTs: p < .05), but not to the HF-meaning targets of
unbalanced homonyms (errors: p = .33; RTs: p = .49).
The results of Experiment 1b were very similar. There was a significant main
effect of Prime [errors: ぬ2(3) = 90.5, p < .001; RTs: ぬ2(3) = 36.9, p < .001]. Planned
contrasts showed less accurate and slower responses to both balanced (both ps <
.001) and unbalanced homonyms (both ps < .001) than to non-homonyms (see
Figure 1 below). Responses were generally less accurate and slower to LF-meaning
than HF-meaning targets, and this main effect of Target [errors: ぬ2(1) = 41.2, p <
.001; ぬ2(1) = 22.0, p < .001] interacted with Prime [errors: ぬ2(3) = 46.4, p < .001; ぬ2(3)
= 39.0, p < .001]. Relative to both targets of non-homonyms, responses were less
accurate and slower to the LF-meaning targets of balanced (both ps < .001) and
unbalanced homonyms (both ps < .001) and the HF-meaning targets of balanced
homonyms (errors: p < .001; RTs: p < .05), as well as less accurate to the HF-
meaning targets of unbalanced homonyms (errors: p < .05; RTs: p = .65).
>> Insert Figure 1 here <<
2.2.2 Unrelated trials
In Experiment 1a, there was only a significant main effect of Prime in error
rates [ぬ2(3) = 10.3, p < .05]. Compared to non-homonyms, responses were less
accurate to balanced homonyms but more accurate to unbalanced homonyms,
AMBIGUITY DISADVANTAGE 16
though neither contrast was significant after the Holm-Bonferroni correction (both ps
= .10). There was also a significant main effect of Prime in RTs [ぬ2(3) = 12.8, p <
.01]. Compared to non-homonyms, responses were slower to balanced (p < .01) but
not unbalanced homonyms (p = .16; see Figure 2 below).
In Experiment 1b, the analyses revealed an unexpected, marginal Prime ×
Target interaction in error rates [ぬ2(3) = 7.7, p = .05] that was due to numerically
higher error rates for the LF-meaning targets of balanced homonyms than one of the
two sets of targets paired with non-homonyms. This contrast, however, was not
significant (p = .14) after the Holm-Bonferroni correction. As in Experiment 1a, there
was a significant main effect of Prime in RTs [ぬ2(3) = 16.9, p < .001]. Compared to
non-homonyms, responses were slower to balanced (p < .001) but not unbalanced
homonyms (p = .10; see Figure 2 below).
>> Insert Figure 2 here <<
2.3 Discussion
Two key findings emerged from Experiment 1. To begin with, we showed that
meaning frequency modulates the ambiguity disadvantage, just like in earlier
investigations into sentence reading (Duffy et al., 1988; Rayner & Duffy, 1986).
Unbalanced homonymy does not produce the disadvantage, most likely due to weak
and delayed activation of the LF meaning in neutral context (Frost & Bentin, 1992;
Klepousniotou et al., 2012; Simpson & Burgess, 1985). This is evident in the finding
that participants rarely selected the LF meaning, even when there was enough time
to do so (high error rates in Experiment 1b), and that a processing cost arose only on
AMBIGUITY DISADVANTAGE 17
trials which instantiated that meaning (higher RTs for unbalanced homonyms than
non-homonyms on LF-meaning trials). It appears, then, that unbalanced homonymy
slows processing only in rare situations when the dominant meaning turns out to be
incorrect, forcing readers to engage in effortful and time-consuming retrieval of the
alternative meaning. We revisit this explanation in Experiment 2.
The pattern of responses was different for balanced homonyms. These words
incurred a significant processing cost whenever they were encountered (higher RTs
for balanced homonyms than non-homonyms on related and unrelated trials). Thus,
not only do we confirm that the ambiguity disadvantage in relatedness decision tasks
is restricted to balanced homonyms, but we also show, for the first time, that this
effect may indeed lie in semantic competition. While the decision-making account
(Pexman et al., 2004) assumes ambiguity to slow processing on related but not
unrelated trials because only the former involves response conflict, our findings
indicate that this is not the case. Experiments 1a and 1b revealed robust ambiguity
disadvantage effects for balanced homonyms even on unrelated trials that are free of
such conflict. We suspect that meaning frequency may be key to explaining the
inconsistencies in findings, especially after discovering that the study by Pexman et
al. (2004) included both balanced and unbalanced homonyms within a single
stimulus list4. It is possible, then, that the study found a null ambiguity effect on
unrelated trials because it did not distinguish between the effects of balanced and
unbalanced homonymy but combined them instead. Given the present evidence
using well-controlled categories of balanced and unbalanced homonymy, the
proposal that the ambiguity disadvantage is due to response-selection difficulties on
related but not unrelated trials appears to lack support, in that it does not
AMBIGUITY DISADVANTAGE 18
accommodate the findings when the role of meaning frequency in the semantic
activation process is taken into account.
Before turning to Experiment 2, it is important to discuss the relatively large
proportion of errors on related trials in Experiments 1a and 1b. While errors were
expected to be very common for homonyms in the LF-meaning condition (Harpaz,
Lavidor, & Goldstein, 2013; Pexman et al., 2004), the results showed that
participants made errors even on trials involving non-homonyms. We think that these
difficulties in detecting and judging the relatedness between primes and targets were
due to multiple constraints in stimulus selection. First, targets were semantic (e.g.,
“tap-sink”) rather than lexical associates5 (e.g., “tap-water”), such that participants
had to retrieve and consider a number of semantic features of the word referents,
which aimed to make the task more sensitive to the impact of semantic activation
(Lucas, 2000; Witzel & Forster, 2014). Second, primes and targets were also
carefully selected and matched on 14 lexical and semantic properties that have been
shown to influence on-line word processing (see Sections 1.1 & 1.2 in the
Supplementary Material).
Certain compromises had to be made as a result of these constraints. In
particular, we note that some of the primes we used may have not been particularly
good at eliciting a given meaning, or as good as they would be when presented
together with an associated particle (e.g., “egg” vs. “egg on” in relation to “urge”;
“tend” vs. “tend to” in relation to “habit”). We were able to address this issue,
however, by demonstrating that high error rates persist and results remain
qualitatively similar when these primes are excluded from analyses (see Section 1.4
in the Supplementary Material). Likewise, we note that some of the targets may have
AMBIGUITY DISADVANTAGE 19
been difficult to process in relation to primes because they had multiple semantically
related senses themselves (e.g., “tend-nurse” where “nurse” could denote a medical
professional, to breast-feed, or to take special care). This was unavoidable given that
over 80% of the words in English are ambiguous in this way (Rodd, Gaskell, &
Marlsen-Wilson, 2002). We did, however, take this into consideration and controlled
for the number of word senses both at the design (Section 1.2 in the Supplementary
Material) and the analysis stage (see Sections 2.2 & 2.3 in the Supplementary
Material). Thus, although our rigorous control over primes and targets may have
contributed to less salient relatedness between the words and less accurate
performance6, we stress that this was instrumental for the design of our study. For
example, matching targets for a large number of control variables, rather than letter
count and/or word frequency alone (Gottlob et al., 1999; Harpaz et al., 2013;
Pexman et al., 2004; Piercey & Joordens, 2000), was necessary to make direct and
reliable comparisons of ambiguity effects in different contexts/prime-target
combinations.
3 Experiment 2
In Experiment 2, we used EEG measurements to establish when the
ambiguity disadvantage for balanced homonyms arises, or at which stage of the
relatedness decision performance. Given that our behavioral results lent support to
the semantic competition account, we expected to observe the ambiguity
disadvantage in the N400 component that has been linked to the ease of semantic
processing. In short, the N400 refers to a negatively-going wave that typically peaks
AMBIGUITY DISADVANTAGE 20
400 ms after the onset of words, pictures, and other meaningful stimuli. Semantic
priming, prior context, and predictability have all been shown to attenuate the relative
amplitude of the N400 to a word, hence the growing consensus is that the
component indexes semantic activation, with larger amplitudes indicating more
effortful form-to-meaning mapping (for a review, see Federmeier & Laszlo, 2009).
Thus, if balanced homonymy produces competition at the semantic level, as seems
to be the case based on our behavioral results, Experiment 2 should show larger
N400 amplitudes for balanced homonyms than non-homonyms. It is critical that this
effect emerges during the reading of the ambiguous prime, separating early
semantic activation processes during prime presentation from late response-
selection processes during target presentation.
Experiment 2 also aimed to further examine the processing of unbalanced
homonyms. Our behavioral results suggest that these words do not produce
semantic competition due to weak and delayed activation of the LF meaning in
minimal context. To substantiate this proposal, we compared the amount of priming
for the HF and LF meanings of balanced and unbalanced homonyms, focusing again
on the N400. The literature on semantic priming has shown that targets preceded by
related primes tend to elicit smaller N400 amplitudes than those preceded by
unrelated primes (for a review, see Kutas & Federmeier, 2011). This “N400 priming”
effect has often been used to investigate patterns of meaning activation in
homonyms, both in isolation (e.g., Atchley & Kwasny, 2003; Klepousniotou et al.,
2012) and biasing context (e.g., Dholakia, Meade, & Coch, 2016; Swaab, Brown, &
Hagoort, 2003). The general finding from such studies is that meanings that are
AMBIGUITY DISADVANTAGE 21
highly frequent or supported by surrounding context are more readily available, and
therefore produce greater N400 priming.
Following this literature, we examined N400 effects to related and unrelated
targets to determine the extent to which the meanings of homonyms are activated
during semantic relatedness decisions. For balanced homonyms, there should be a
comparable N400 priming effect for targets instantiating either of the meanings. This
would indicate that both meanings are activated to the same extent and in parallel.
For unbalanced homonyms, on the other hand, there should be substantial priming
for the HF meaning, but little or even no priming for the LF counterpart. This would
support our idea that, in isolation or neutral context, readers typically fail to
comprehend unbalanced homonyms in the unexpected alternative meaning due to
reduced and insufficient activation of that meaning.
3.1 Method
3.1.1 Participants
A different group of 34 University of Leeds students and staff [27 females,
aged 18-33 (M = 20.9, SD = 3.5)] participated in exchange for course credit or £8. All
participants were right-handed monolingual native British-English speakers with no
known history of any language-/vision-related difficulties or neurological damage or
disorders. The experiment received ethical approval from the School of Psychology,
University of Leeds Ethics Committee.
AMBIGUITY DISADVANTAGE 22
3.1.2 Stimuli & procedure
Experiment 2 involved the same task and stimuli as Experiment 1b, but there
were four minor changes to the procedure. First, participants responded with a
computer mouse, rather than a keyboard. Second, there were four, rather than two,
one-minute breaks – one before each experimental block. Third, we used a longer
inter-trial interval (1000 ms instead of 100 ms) to allow participants to blink and rest
their eyes, and there was a 200 ms interval between the target and response
execution that aimed to minimize any overlap in ERP components evoked by these
trial events. Trials began with a 500 ms fixation cross. After 100 ms, the prime and
the target were presented for 700 ms and 500 ms, respectively, with a 50 ms interval
in between. Once the target disappeared, there was a 200 ms interval followed by a
1500 ms visual cue (“???”) for response execution. Trials ended with a 1000 ms
inter-trial-interval (ITI). Fourth, instructions and feedback within the practice block
emphasized response accuracy only. Effects in RTs were of no particular interest
because Experiment 2 involved a delayed response paradigm, which may have to
some extent contaminated our measure of lexical-semantic processing.
3.1.3 EEG data acquisition
The EEG was recorded using 64 pin-type active Ag/AgCl electrodes that were
embedded in a head cap, arranged according to the extended 10-20 positioning
system (Sharbrough et al., 1991), and connected to a BioSemi ActiveTwo AD-box
with an output impedance of less than 1っ (BioSemi, Amsterdam, the Netherlands).
Recording involved 10 midline electrodes and 27 electrodes placed over each
AMBIGUITY DISADVANTAGE 23
hemisphere. Ground electrodes were placed between Cz and CPz. Eye movements
were monitored using four electrodes – bipolar horizontal electro-oculogram (EOG)
was recorded between the outer right and left canthi, and bipolar vertical EOG was
recorded above and below the left eye. Additional electrodes were placed on the left
and the right mastoid. The EEG and EOG were recorded continuously with a
bandpass filter of 0.16-100 Hz and digitised at a 512-Hz sampling rate.
3.1.4 EEG data pre-processing
The EEG was pre-processed off-line using MATLAB (The Mathworks, Natick,
Massachusetts) and EEGLAB (Delorme & Makeig, 2004). The data were first down-
sampled to 250 Hz, referenced to the algebraic average of the left and the right
mastoid, and then filtered (0.1 - 40 Hz, 12 dB/Oct, Butterworth zero phase filter).
Blinks, eye movements, muscle activity, bad channels, and other artifacts were
corrected for based on independent component analysis (ICA) guided by measures
from SASICA (Chaumon, Bishop, & Busch, 2015; on average, 2-4 components per
participant were removed). Cleaned data were then segmented into two types of
epochs. For prime-window analyses, epochs started 100 ms before and ended 700
ms after the onset of the prime. For target-window analyses, epochs started 50 ms
before the onset of the target and ended 200 ms after the offset. The 100/50 ms
intervals before the onset of the prime/target were used to normalize the onset
voltage of the ERP waveform.
3.2 Results
AMBIGUITY DISADVANTAGE 24
3.2.1 Behavioural data
Two of the 34 participants were removed from all analyses – one due to a
relatively large number of errors on related (37.1%) and unrelated trials (25.9%) and
the other due to a large number of epochs containing amplifier saturation artifacts
(+/- 100 µV; 49.0% in the prime window, 54.6% in the target window). Accuracy and
latency data were analyzed in the same way as in Experiment 1. For RTs, we
excluded errors (13.2% of trials) and any responses that were two SDs above/below
a participant’s mean in a given condition (4.9% of trials). We did not transform the
remaining RTs as the residuals from linear mixed-effects models followed a normal
distribution. All models included Prime and Target as fixed effects as well as random
intercepts for subjects and items. As in Experiment 1, RT models included Block as
an additional fixed effect.
The results were similar to those of Experiments 1a and 1b. For related trials,
there was a significant main effect of Prime in both error rates [ぬ2(3) = 56.7, p < .001]
and RTs [ぬ2(3) = 34.2, p < .001]. Planned contrasts showed less accurate and slower
responses to both balanced (both ps < .001) and unbalanced homonyms (both ps <
.001) than to non-homonyms (see Figure 3 below). Responses were generally less
accurate and slower to LF-meaning than HF-meaning targets, and this main effect of
Target [errors: ぬ2(1) = 32.0, p < .001; RTs: ぬ2(1) = 23.8, p < .001] interacted with
Prime [errors: ぬ2(3) = 49.3, p < .001; RTs: ぬ2(3) = 40.5, p < .001]. Relative to both
targets of non-homonyms, responses were less accurate and slower to the LF-
meaning targets of balanced (both ps < .001) and unbalanced homonyms (both ps <
.001), but not to the HF-meaning counterparts of balanced (errors: p = .08; RTs: p =
.08) or unbalanced homonyms (errors: p = .56; RTs: p = .44). For unrelated trials,
AMBIGUITY DISADVANTAGE 25
there was only a significant main effect of Prime in RTs [ぬ2(3) = 24.6, p < .001].
Compared to non-homonyms, responses were slower to balanced (p < .01) but not
unbalanced homonyms (p = .10).
>> Insert Figure 3 here <<
Before turning to EEG data, it is important to note that although Experiments 1
and 2 showed the same patterns of responses for balanced and unbalanced
homonyms, the overall error rates appeared to be lower in Experiment 2 (compare
Figures 1-3). In order to examine this further, we decided to contrast participants’
performance in Experiments 1b and 2 as these were the most similar with respect to
stimulus-presentation procedures. The analyses below were the same as those
conducted for each experiment separately, except that they included the factor of
Experiment (in addition to Prime and Target). All models included significant random
intercepts for subjects and items as well as a random slope for Experiment across
subjects.
For related trials, there was a significant main effect of Experiment [ぬ2(1) =
7.5, p < .01], with higher error rates in Experiment 1b (M = 30.4, SD = 8.9) than
Experiment 2 (M = 24.0, SD = 6.2). There was also a significant Experiment × Prime
interaction [ぬ2(3) = 26.8, p < .001]. Post hoc tests indicated that the simple effect of
Experiment (i.e., higher error rates in Experiment 1b) was significant for balanced
(Experiment 1b: M = 30.5, SD = 12.4; Experiment 2: M = 27.3, SD = 8.2; p < .001)
and unbalanced homonyms (Experiment 1b: M = 50.1, SD = 11.2; Experiment 2: M =
42.7, SD = 8.9; p < .01), but not for non-homonyms (Experiment 1b: M = 15.6, SD =
8.1; Experiment 2: M = 13.0, SD = 6.8; p = .35). The simple effect of Experiment was
AMBIGUITY DISADVANTAGE 26
also significantly greater for balanced than unbalanced homonyms (p < .001). For
unrelated trials, there was only a significant main effect of Experiment [ぬ2(1) = 7.7, p
< .01], with higher error rates in Experiment 1b (M = 11.9, SD = 2.7) than Experiment
2 (M = 10.4, SD = 2.9).
These results suggest that detecting and judging semantic relatedness was
somewhat easier in Experiment 2, especially on trials involving ambiguous words.
One particularly important difference between the experiments concerned the
instructions given to participants and feedback within the practice block. While in
Experiment 1 the instructions and training emphasized both response speed and
accuracy, in Experiment 2 they emphasized accuracy only (RTs were of no particular
interest due to the delayed response paradigm in the experiment). We think that not
only does this explain why accuracy was superior in Experiment 2, but it also sheds
some light on our task in general. It appears that the fast-paced nature of our task, or
over-emphasis on speed on participants’ part, may have to some extent
compromised accuracy. This, coupled with the use of more difficult prime-target word
pairs, as discussed earlier, could explain why error rates in the present study were
relatively high even in the easier conditions involving homonyms in the HF meaning
and non-homonyms.
3.2.2 EEG data
EEG analyses excluded individual epochs containing amplifier saturation
artifacts (+/- 100 µV; 0.9% of trials in the prime window, 1.3% in the target window)
or errors (12.5% of all trials). Following recent studies (Amsel, 2011; De Cat,
Klepousniotou, & Baayen, 2015; Kornrumpf, Niefind, Sommer, & Dimigen, 2016),
AMBIGUITY DISADVANTAGE 27
epoched data were analysed on a trial-by-trial basis using linear mixed-effects
modelling, primarily due to a large number of errors on LF-meaning trials. As in De
Cat et al. (2015), we analyzed each of the 64 channels separately as there was too
much data (over 700,000 observations per channel) to fit a single model. In order to
prevent spurious results due to a potential multiplicity problem, we used
topographical consistency as an additional criterion when judging the reliability of
results. The rationale was that since channels are not entirely independent, any
effects specific to ambiguity should be similar across neighbouring channels.
Since our hypotheses for both the prime and the target window concerned the
N400, analyses focused on the 350-500 ms segment, which best represented this
component in our data. Visual inspection of the waveforms within the segment during
prime (see Figure 4 below) and target presentation (see Figures 5 & 6 below)
revealed a large difference in peak latency for unbalanced homonyms during target
presentation (i.e., much earlier peaks for the HF-meaning than LF-
meaning/unrelated targets). Thus, as in previous ERP studies of semantic influences
on reading (e.g., Taler, Kousaie, & Lopez Zunini, 2013), we divided the 350-500 ms
segment into four consecutive time bins of 50 ms in order to capture and account for
the divergence in the waveforms.
3.2.2.1 Prime presentation
Prime-window analyses compared N400 amplitudes to homonymous and
non-homonymous words during prime presentation. This involved a set of mixed-
effects models with the factors of Prime (balanced homonym, unbalanced homonym,
non-homonym1, non-homonym2), Time (350-400 ms, 400-450 ms, 450-500 ms, 500-
AMBIGUITY DISADVANTAGE 28
550 ms), and Block (1, 2, 3, 4). Models included random intercepts for subjects and
items as well as random by-subject slopes (mainly for Time). Planned contrasts
compared balanced/unbalanced homonyms to both sets of non-homonyms, and their
significance level was adjusted using the Holm-Bonferroni method. Only significant
effects that involved Prime and were relevant to the hypotheses are reported below.
There was a significant Prime × Time interaction at all channels (all ps < .05),
except for T7, TP7, and P9 (for full test results for each channel and effect, see
Section 3 in the Supplementary Material). Amplitudes in the 400-450 ms window
were larger (i.e., more negative) for balanced homonyms than non-homonyms at
fronto-polar (FPz), anterio-frontal (AFz, AF3, AF4, AF8), frontal (Fz, F1, F3, F2, F4),
fronto-central (FCz, FC1, FC3, FC2), and fronto-temporal sites (FT7, FT8). This
effect also occurred at similar sites in the earlier 350-400 ms (FPz, AF8, Fz, F2, F4,
FT7, & FT8; all ps < .05) and the later 450-500 ms window(FPz, AF3, AF4, AF8, Fz,
F1, F2, F4, FT7, & FT8; all ps < .05). There were no significant differences between
unbalanced homonyms and non-homonyms (see Figure 4 below). Overall, then, the
prime-window analyses showed increased negativity from 350 ms to 500 ms post-
prime onset for balanced but not unbalanced homonyms. This effect appeared over
bilateral medial frontal sites, extending anteriorly to anterio-frontal sites and
posteriorly to fronto-temporal sites.
>> Insert Figure 4 here <<
3.2.2.2 Target presentation
AMBIGUITY DISADVANTAGE 29
Target-window analyses compared N400 amplitudes to the related and
unrelated targets of homonyms. This involved a set of mixed-effects models with the
factors of Prime (balanced homonym, unbalanced homonym), Target (HF-meaning,
LF-meaning, unrelatedA, unrelatedB), Time (350-400 ms, 400-450 ms, 450-500 ms,
500-550 ms), and Block (1, 2, 3, 4). Non-homonyms were excluded as the aim was
to examine the amount of priming for the different meanings of balanced versus
unbalanced homonyms. Models included random intercepts for subjects and items
and random by-subjects slopes (mainly for Block or Target). Planned contrasts
compared HF- and LF-meaning targets to each other and to both sets of unrelated
targets, and their significance threshold was adjusted using the Holm-Bonferroni
method. Only significant effects that involved Target and were relevant to the
hypotheses are reported below.
There was a significant main effect of Target (all ps < .05) at fronto-central
(FCz, FC1, FC2), central (Cz, C1, C3, C5, C2, C4), centro-parietal (CPz, CP1, CP3,
CP5, CP2, CP4, CP6), parietal (Pz, P1, P3, P5, P2, P4, P6, P8), parieto-occipital
(POz, PO3, PO4, PO8), and occipital sites (Oz, O1, O2). Planned contrasts showed
reduced (i.e., less negative) amplitudes to HF-meaning targets (all ps < .05) relative
to unrelated and LF-meaning targets at most of these channels. There were no
significant differences between LF-meaning and unrelated targets.
There was a significant Target × Time interaction (all ps < .05) at all channels,
except for P9. HF-meaning targets elicited smaller amplitudes than unrelated targets
in the 400-450 ms, 450-500 ms, and 500-550 ms windows at all the channels (all ps
< .05), except for AF7, AF8, F5, F7, F8, FT7, T7, TP7, P7, and P10. In addition, HF-
meaning targets elicited smaller amplitudes than LF-meanings targets in the 500-550
AMBIGUITY DISADVANTAGE 30
ms window at frontal (Fz, F2), fronto-central (FCz, FC1, FC2, FC4), central (Cz, C1,
C3, C5, C2, C4), centro-parietal (CP1, CP3, CP5, CP2, CP4, CP6), parietal (Pz, P1,
P3, P5, P2, P4, P6), parieto-occipital (POz, PO3, PO4), and occipital sites (Oz, O1,
O2), as well as the inion (Iz; all ps < .05). This effect was also significant in the
earlier 400-450 ms and 450-500 ms windows at the same channels (all ps < .05),
except for F2, FC4, C1, C5, C4, CP3, CP5, CP6, P2, and Iz. LF-meaning targets, on
the other hand, elicited smaller amplitudes than unrelated targets at CP4, CP6, P2,
POz, and PO8 in the 450-500 ms window only (all ps < .05).
There was a significant Target × Time × Prime interaction (all ps < .05) at all
channels, except for T7, TP7, P7, and P9. For balanced homonyms (see Figure 5
below), HF-meaning targets elicited smaller amplitudes than unrelated targets in the
last 500-550 ms window at fronto-central (FCz, FC1, FC3, FC2, FC4), central (Cz,
C1, C3, C5, C2, C4), centro-parietal (CPz, CP1, CP3, CP5, CP2, CP4, CP6), parietal
(Pz, P1, P3, P5, P2, P4, P6, P8), and parieto-occipital sites (POz, PO3, PO4; all ps <
.05). This effect also occurred in the earlier 450-500 ms window at a smaller cluster
(Cz, CPz, CP1, CP2, CP6, Pz, P1, P2, POz, PO3, & PO4; all ps < .05). The
contrasts between the HF-meaning and LF-meaning targets as well as between the
LF-meaning and unrelated targets for balanced homonyms were not significant.
>> Insert Figure 5 here <<
For unbalanced homonyms (see Figure 6 below), on the other hand, HF-
meaning targets elicited smaller amplitudes than unrelated targets in the 450-500 ms
and 500-550 ms windows at all the channels (all ps < .05), except for AF7, AF8, F5,
AMBIGUITY DISADVANTAGE 31
F7, FT7, and P10 (in addition to the four channels that did not show the 3-way
interaction in the first place). This effect also occurred in the earlier 400-450 ms
window (all ps < .05) at the same channels, except for AF3, FP1, F3, FC5, TP8,
CP5, P8, PO7, and Iz. In addition, HF-meaning targets elicited smaller amplitudes
than LF-meaning targets in the last 500-550 ms window at fronto-polar (FPz, FP1,
FP2), anterio-frontal (AFz, AF3, AF4), frontal (Fz, F1, F2, F4, F6), fronto-central
(FCz, FC1, FC3, FC2, FC4), central (Cz, C1, C3, C2, C4, C6), centro-parietal (CPz,
CP1, CP2), parietal (Pz, P1, P3, P2, P4), parieto-occipital (POz, PO3, PO4), and
occipital sites (Oz, O1, & O2; all ps < .05). This effect also occurred at similar sites in
the earlier 400-450 ms (FPz, Fz, F4, F6, FCz, FC1, FC2, Cz, C1, C2, C6, CPz, CP1,
CP2, Pz, P1, P3, PO3, Oz, & O2) and 450-500 ms windows (AFz, AF3, AF4, FPz,
FP1, FP2, Fz, F1, F4, F6, FCz, FC1, FC3, FC2, FC4, Cz, C1, C2, C4, C6, CPz, CP1,
CP2, Pz, P1, P3, PO3, Oz, & O2; all ps < .05). The contrasts between the LF
meanings and unrelated targets for unbalanced homonyms were not significant.
In summary, the target-window analyses showed that amplitudes to the HF-
meaning targets of balanced homonyms were reduced only in comparison to
unrelated targets, primarily from 500 ms to 550 ms post-target onset. Amplitudes to
the HF-meaning targets of the unbalanced counterparts, on the other hand, were
reduced in comparison to both unrelated and LF-meaning targets, and this effect
was markedly sustained (400-550 ms post-target onset). For both balanced and
unbalanced homonyms, priming for the HF meaning appeared over bilateral medial
and lateral centro-parietal sites, extending anteriorly to frontal sites and posteriorly to
occipital sites. In contrast, no significant differences were observed between LF-
meaning and unrelated targets for either balanced or unbalanced homonyms.
AMBIGUITY DISADVANTAGE 32
>> Insert Figure 6 here <<
3.3 Discussion
Two key findings emerged from Experiment 2. To begin with, analyses for the
prime window revealed increased frontal negativity from 350 ms to 500 ms post-
prime-onset for balanced but not unbalanced homonyms (relative to non-
homonyms), which, as we argue in the section below, is compatible with the
semantic competition account (Armstrong & Plaut, 2008; Kawamoto, 1993; Rodd et
al., 2004). The finding that homonymy in general had an impact on the N400 is
consistent with previous lexical decision studies which reported greater N400
responses to homonyms than non-homonyms (Haro, Demestre, Boada, & Ferré,
2017; see also Beretta, Fiorentino, & Poeppel, 2005; MacGregor et al., 2020). Not
only does our experiment corroborate and extend this work by demonstrating that
homonymy also affects the N400 component in semantically engaging tasks that
require disambiguation, but it also shows that it is balanced, not unbalanced,
homonymy that drives this effect. In other words, our study is the first to provide EEG
evidence for the long-held assumption that meaning frequency modulates ambiguity
effects in word processing (for behavioral evidence, see Experiment 1; Armstrong et
al., 2012; Brocher et al., 2018; Rayner & Duffy, 1986).
Turning to the analyses for the target window, the results confirmed that
balanced and unbalanced homonyms differ in the extent to which their meanings are
activated in the absence of context. There was a significant N400 priming effect for
HF-meaning targets and a non-significant one for LF-meaning targets (relative to
AMBIGUITY DISADVANTAGE 33
unrelated targets), both for balanced and unbalanced homonyms. Note, however,
that there was evidence to suggest that (weaker) priming also occurred for the LF
meaning of balanced homonyms. Targets instantiating that meaning elicited N400
amplitudes that were (a) numerically, though not statistically, smaller than those for
unrelated targets and (b) comparable to those for HF-meaning targets (see Figure 5
above). In other words, while the dominant meaning was activated and facilitated the
processing of the related target for both types of homonyms, the alternative
counterpart was activated (to a lesser degree) only for balanced homonyms. This
suggests that balanced and unbalanced homonyms differ in how and when their
meanings are activated, and may therefore produce different levels of semantic
competition.
4 General Discussion
The present study provides consistent behavioral and electrophysiological
evidence that the ambiguity disadvantage is due to competition between multiple
semantic representations during the activation process, as predicted by current
connectionist models of ambiguity (Armstrong & Plaut, 2008; Kawamoto, 1993; Rodd
et al., 2004). Experiment 1 shows that the ambiguity disadvantage arises for
balanced but not unbalanced homonyms, extending previous findings from sentence
reading (Duffy et al., 1988; Rayner & Duffy, 1986) to single-word processing. This
effect was not restricted to related trials, which proves particularly challenging for the
decision-making account proposed by Pexman et al. (2004). While the account
assumes ambiguity to slow relatedness decisions on related but not unrelated trials
AMBIGUITY DISADVANTAGE 34
because only the former involves response conflict, we demonstrate that this is not
the case - balanced homonymy incurred a processing cost regardless of whether the
different meanings triggered consistent or inconsistent responses to the target (i.e.,
both on unrelated and related trials). This is in line with our recent finding that
learning new meanings for familiar words slows the processing of their existing
meanings (mirroring the ambiguity disadvantage in natural language), both on
related and unrelated trials in relatedness decision tasks (Maciejewski, Rodd, Mon-
Williams, & Klepousniotou, 2019). Overall, then, it appears that the idea that the
ambiguity disadvantage is due to additional decision making involved in response-
conflict resolution faces a major challenge, in that it cannot explain why balanced
homonymy would incur a processing cost on unrelated trials that are free of
response conflict.
Further evidence against the decision-making account comes from
Experiment 2. To begin with, the finding that the effect of balanced homonymy was
observed during prime presentation confirms that the ambiguity disadvantage arises
when processing the ambiguous word itself, rather than when processing or
responding to the target. More specifically, it arises during the semantic activation
process, as revealed by increased negativity in the N400 window. Note, however,
that while the latency of our effect is consistent with that of a typical N400 effect, this
is not the case with respect to scalp topography. The ERP literature (for a review,
see Kutas & Federmeier, 2011) shows that the “traditional” N400 effect is normally
largest over centro-parietal sites, rather than frontal sites as in the current study,
though there have been reports of increased frontal negativity for homonyms versus
non-homonyms before (Lee & Federmeier, 2006, 2009; see also Mollo, Jefferies,
AMBIGUITY DISADVANTAGE 35
Cornelissen, & Gennari, 2018). This striking difference in topography suggests that
the common explanation for an N400 effect in terms of differences in the extent of
semantic activation or priming may not fully apply to our effect. Increased frontal
negativity for balanced homonymy may instead point to an additional, inhibitory
process involved in ambiguity resolution – most likely semantic competition, as
suggested by the literature reviewed next.
fMRI studies of ambiguity found that the left inferior frontal gyrus (LIFG), in
particular pars triangularis (BA 45) and pars opercularis (BA 44), is the most
consistent brain region to show an increased haemodynamic response to ambiguity
(for a detailed review, see Vitello & Rodd, 2015), though there is also evidence for
bilateral recruitment of that area when processing ambiguous words (Klepousniotou,
Gracco & Pike, 2014). There also appears to be wide agreement in this literature that
the LIFG is involved in the resolution of competition between the multiple meanings
of an ambiguous word, either when the word is encountered in isolation (e.g.,
Bilenko, Grindrod, Myers, & Blumstein, 2009; Hargreaves et al., 2011) or when the
word must be reinterpreted following initial selection of the incorrect meaning (e.g.,
Rodd, Johnsrude, & Davis, 2012). The former situation closely corresponds to the
prime window in Experiment 2, hence increased frontal negativity for balanced
homonyms in that window may indicate competition between their meanings within
the LIFG. This interpretation is further supported by the influential “conflict resolution”
account of LIFG function (e.g., Novick, Trueswell, & Thompson-Schill, 2009;
Thompson-Schill, D’Esposito, Aguirre, & Farah, 1997), according to which posterior
LIFG serves to resolve competition between multiple representations. In particular,
Novick et al. (2009) proposed that posterior LIFG engages in the resolution of
AMBIGUITY DISADVANTAGE 36
competition, regardless of its specific linguistic form, either when there is a prepotent
but irrelevant response, or when there are multiple activated representations but no
dominant response. Since reading balanced homonyms in Experiment 2 produced
the latter type of competition, at the semantic level in this case, increased frontal
negativity for these words may indeed reflect increased activation of the LIFG, and
its RH homologue, in an attempt to resolve that competition7.
Overall, then, the present findings are incompatible with the idea that
response conflict constitutes an explanation for the ambiguity disadvantage (Pexman
et al., 2004). In particular, the finding that balanced homonymy affected the N400
component in the prime window indicates that the effect arises during the semantic
activation of the ambiguous prime, hundreds of milliseconds before participants see
the related/unrelated target that follows. This clearly shows that the ambiguity
disadvantage is not due to response-selection difficulties upon target presentation,
but due to semantic competition in response to ambiguity itself. Note, however, that
the present findings are not necessarily incompatible with Hino et al.’s (2006)
decision-making account that focuses on qualitative task differences and their impact
on how the response system is configured, rather than response conflict. Under this
account, the ambiguity disadvantage arises only when a task places demands on
post-semantic processes, such as analysis of semantic features, that support
response selection. We do not provide compelling evidence either for or against this
account, since our study aimed to examine the ambiguity disadvantage in semantic
relatedness decisions, rather than semantic categorisations that the account focuses
on. We do, however, think that it is possible to marry some aspects of the decision-
making account with the semantic competition one. In particular, we agree with Hino
AMBIGUITY DISADVANTAGE 37
et al. (2006) that the impact of ambiguity in word processing largely depends on what
readers/listeners must do with the word. For instance, in relatedness decision tasks,
competition effects arise for homonyms because responses are made based on
complete semantic activation, in the sense that participants must settle on a
particular meaning of these words. In lexical decision tasks, competition effects
become noticeable only when responses are more reliant on semantic activation,
when, for example, discriminating between homonyms and pseudo-homophonic
(Azuma & Van Orden, 1997; Rodd et al., 2002) or wordlike non-words (Armstrong &
Plaut, 2008, 2016). Likewise, in semantic categorization tasks, competition effects
arise only when there is a need for greater semantic activation, when responses
cannot be made based on a small number of semantic features (Hino et al., 2006).
Thus, the general idea is that task demands play some role in generating ambiguity
effects. However, while Hino et al. (2006) assert that this is due to differences in how
the response system is configured in a particular task, we suggest that this is more
likely due to differences in the level of semantic activation needed to perform the
task (for a similar view, see Armstrong & Plaut, 2016). This is supported by our
demonstrations that meaning frequency, which influences the level of semantic
activation, modulates the ambiguity disadvantage, and that the ambiguity
disadvantage arises during semantic, rather than post-semantic, processes.
In contrast, the present findings are readily compatible with the predictions of
current connectionist models of ambiguity (Armstrong & Plaut, 2008; Kawamoto,
1993; Rodd et al., 2004). Explaining why meaning frequency would modulate the
ambiguity disadvantage presents these models with little challenge. Within the
connectionist framework, long-term experience with a particular meaning of an
AMBIGUITY DISADVANTAGE 38
ambiguous word modifies the strength of the connections between orthographic and
semantic units, which in turn determines the speed and outcome of form-to-meaning
mapping. The HF meanings of unbalanced homonyms develop strong connections,
and thus are activated so fast that they avoid competition with the LF meanings. For
balanced homonyms, meaning frequency plays barely any role in form-to-meaning
mapping - both meanings are activated to the same extent and in parallel, with each
being equally likely to win competition for further activation. Therefore, connectionist
models of ambiguity, such as the one implemented by Kawamoto (1993), can easily
account for the differential effects of balanced and unbalanced homonymy in
semantic activation (and competition involved) by modifying the weights on the
connections between orthographic and semantic units.
Note that these differential effects are evident in both our behavioral and
electrophysiological data. For unbalanced homonyms, activation of the LF meaning
was so weak that participants rarely selected that meaning in minimal context, even
when there was enough time do so. This is in line with the finding of very high error
rates for unbalanced homonyms in the LF meaning in the short (Experiment 1a) and
the long prime-duration condition (Experiments 1b & 2). When participants did
disambiguate the words towards the LF meaning, there was a substantial processing
cost (higher RTs for unbalanced homonyms than non-homonyms on correct LF-
meaning trials in Experiments 1 & 2) that we take as evidence of effortful and slow
retrieval of that meaning upon seeing a supporting target. This is in line with the
finding of no N400 priming for the LF-meaning target even on correct trials
(Experiment 2) as well as the finding of a similar processing cost at unexpected LF-
meaning context following an unbalanced homonym (e.g., “We knew the boxer was
AMBIGUITY DISADVANTAGE 39
barking all night”) in eye-movement studies (Duffy et al., 1988; Rayner & Duffy,
1986).
Importantly, the difficulty in processing the LF meaning of unbalanced
homonyms did not arise because participants did not know that meaning.
Maciejewski and Klepousniotou’s (2016) norms, from which the homonyms were
selected, confirm that over 75 out of the 100 native speakers they tested used and/or
encountered the LF meaning of these words8. This suggests that, for most
participants in the current study, the meaning was stored in the mental lexicon but
not sufficiently activated in the absence of context. Support for this interpretation
comes from the finding that readers struggle but eventually manage to understand
unbalanced homonyms in the LF meaning solely based on strong sentential context
(e.g., Brocher et al., 2018; Duffy et al., 1988; Leinenger, Myslín, Rayner, & Levy,
2017). Further support comes from stimulus pre-tests that we conducted as part of
the study (see Section 1.3 in the Supplementary Material). These pre-tests showed
that raters normally failed to detect the semantic relatedness between unbalanced
homonyms and targets instantiating the LF meaning, unless they were first
presented with sentential context supporting that meaning. The implication is that
naturalistic and elaborate context may be necessary to fully retrieve and select the
LF counterpart, both in on-line and off-line tasks. This is because, for ease of
comprehension, the language system appears to process unbalanced homonyms as
functionally unambiguous words.
For balanced homonyms, the results indicate that although both their
meanings were sufficiently activated to produce semantic competition, they did not
seem to be activated to the same extent. After all, there were fewer errors
AMBIGUITY DISADVANTAGE 40
(Experiments 1 & 2), faster responses (Experiments 1 & 2), and larger N400 priming
(Experiment 2) on HF-meaning than LF-meaning trials. This should not come as a
surprise. Truly balanced homonyms are very rare at best (see Armstrong et al.,
2012), hence the relative frequencies of the meanings of our words differed, on
average, by 20% (SD = 12). It appears, then, that even balanced homonyms show
small, albeit noticeable, bias in the activation process.
Lastly, we wish to emphasize that although our study lends support to
the semantic competition account, it does not really help to distinguish between
specific connectionist models of ambiguity that proposed the account (Armstrong &
Plaut, 2008; Kawamoto, 1993; Rodd et al., 2004). While all three models predict
homonymy to produce semantic competition in tasks that require meaning selection
(e.g., semantic relatedness decisions), they disagree quite substantially on the
impact of homonymy in tasks that do not (e.g., lexical decisions). Kawamoto’s (1993)
model predicts a facilitatory effect due to enhanced feedback from semantics during
orthographic processing, which is at odd with most lexical decision studies (for a
review, see Eddington & Tokowicz, 2015). Rodd et al.’s (2004) model, on the other
hand, predicts an inhibitory effect due to inconsistent form-to-meaning mappings
during semantic processing, regardless of task demands. Armstrong and Plaut’s
(2008) model also predicts an inhibitory effect, but only when the task is sufficiently
difficult to engage substantial semantic processing. Not only do the models disagree
on why and when ambiguity has its effect, but they also differ in terms of descriptions
of the roles of context, meaning frequency, and meaning relatedness. Kawamoto’s
(1993) model simulates the predicted effects of meaning frequency but does not
make the important distinction between homonyms and polysemes (i.e., words with
AMBIGUITY DISADVANTAGE 41
multiple related senses, such as “nurse”). Rodd et al.’s (2004) and Armstrong and
Plaut’s (2008) models make the distinction, but only the latter discusses (but does
not simulate) the roles of context and meaning frequency. Taken together, while our
study supports the overall semantic competition account, more evidence is needed
to advance or constrain the models. In particular, future studies should attempt to
extend our findings to other forms of ambiguity, given growing evidence that for
polysemes semantic competition may largely depend on the degree of overlap of the
multiple senses (Windisch Brown, 2008; Klepousniotou, Titone, & Romero, 2008;
Maciejewski et al., 2019), such that it could be minimal for words with highly
overlapping senses (e.g., “dust”) but strong, albeit not as much as for homonyms, for
words with less overlapping senses (e.g., “virus”).In conclusion, the present findings
demonstrate that the ambiguity disadvantage in relatedness decision tasks is
restricted to balanced homonyms and show, for the first time, that this effect arises
during the semantic processing of the ambiguous word itself. More specifically, the
study suggests that balanced homonyms give rise to competition during the
semantic activation process which most likely engages the LIFG that has been
implicated in the resolution of such competition (for a review, see Vitello & Rodd,
2015). In addition, the study provides direct evidence that balanced and unbalanced
homonyms differ in how their meanings are activated out of context, which
determines the degree of competition they produce.
The present findings are consistent with semantic competition accounts
proposed by connectionist models of ambiguity, especially those that incorporate an
explanation for the role of meaning frequency (Kawamoto, 1993). They are not,
however, consistent with decision-making accounts, especially those that attribute
AMBIGUITY DISADVANTAGE 42
the ambiguity disadvantage to response-conflict resolution (Pexman et al., 2004). In
particular, such accounts fail to accommodate the finding of the ambiguity
disadvantage during the semantic processing of the ambiguous word itself, rather
than during the processing of the related/unrelated target and subsequent response
making. Furthermore, if the ambiguity disadvantage is merely a task artifact at the
response-selection stage, it remains unclear why it would be robust across a number
of tasks of distinct response-selection demands. After all, competition effects in
ambiguity resolution have been observed in tasks involving semantic relatedness
(e.g., Gottlob et al., 1999) and categorisation decisions (e.g., Jager & Cleland, 2015),
semantically primed (e.g., Klepousniotou, 2002) and unprimed lexical decisions (e.g.,
Armstrong & Plaut, 2016), sensicality judgements (e.g., Klepousniotou et al., 2008),
and even sentence-reading tasks that do not require any response or decision (e.g.,
Duffy et al., 1988). Our study marks a significant step towards unravelling the locus
of these competition effects, in that it establishes that, at least in relatedness
decision tasks, these effects arise due to semantic activation processes.
AMBIGUITY DISADVANTAGE 43
Acknowledgements
This work was funded by an Economic and Social Research Council doctoral
studentship (ES/J500215/1) and a University of Leeds postgraduate research grant
awarded to the first author. We thank Rebecca Gardner, Marketa Provodova, and
Timothy Carling for their help with data collection, and Brian Scully for his help with
data pre-processing. We also thank three anonymous reviewers for their helpful
comments on an earlier draft of the manuscript.
AMBIGUITY DISADVANTAGE 44
References
Amsel, B. D. (2011). Tracking real-time activation of conceptual knowledge using
single-trial event-related potentials. Neuropsychologia, 49(5), 970-983.
Annett, M. (1967). The binomial distribution of right, mixed and left handedness. The
Quarterly Journal of Experimental Psychology, 19(4), 327-333.
Armstrong, B. C., & Plaut, D. C. (2008). Settling dynamics in distributed networks
explain task differences in semantic ambiguity effects: Computational and
behavioral evidence. In B. C. Love, K. McRae, & V. M. Sloutsky (Eds.),
Proceedings of the 30th Annual Conference of the Cognitive Science
Society (pp. 273-278). Austin, TX: Cognitive Science Society.
Armstrong, B. C., & Plaut, D. C. (2016). Disparate semantic ambiguity effects from
semantic processing dynamics rather than qualitative task differences.
Language, Cognition, and Neuroscience, 1(7), 1-27.
Armstrong, B. C., Tokowicz, N., & Plaut, D. C. (2012). eDom: Norming software and
relative meaning frequencies for 544 English homonyms. Behavior Research
Methods, 44(4), 1015-1027.
Atchley, R. A., & Kwasny, K. M. (2003). Using event-related potentials to examine
hemispheric differences in semantic processing. Brain and Cognition, 53(2),
133-138.
Azuma, T., & Van Orden, G. C. (1997). Why SAFE is better than FAST: The
relatedness of a word's meanings affects lexical decision times. Journal of
Memory and Language, 36(4), 484-504.
AMBIGUITY DISADVANTAGE 45
Barr, D. J., Levy, R., Scheepers, C., & Tily, H. J. (2013). Random effects structure
for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and
Language, 68(3), 255-278.
Bates, D., Mächler, M., & Bolker, B., (2011). Lme4: Linear mixed-effects modeling
using S4 classes [R package]. Retrieved from http://CRAN.R-Project.org/
package=lme4.
Beretta, A., Fiorentino, R., & Poeppel, D. (2005). The effects of homonymy and
polysemy on lexical access: An MEG study. Cognitive Brain Research, 24(1),
57-65.
Bilenko, N. Y., Grindrod, C. M., Myers, E. B., & Blumstein, S. E. (2009). Neural
correlates of semantic competition during processing of ambiguous
words. Journal of Cognitive Neuroscience, 21(5), 960-975.
Briggs, G. G., & Nebes, R. D. (1975). Patterns of hand preference in a student
population. Cortex, 11(3), 230-238.
Brocher, A., Koenig, J. P., Mauner, G., & Foraker, S. (2018). About sharing and
commitment: The retrieval of biased and balanced irregular
polysemes. Language, Cognition and Neuroscience, 33(4), 443-466.
Chaumon, M., Bishop, D. V. M., & Busch, N. A. (2015). A practical guide to the
selection of independent components of the electroencephalogram for artifact
correction. Journal of Neuroscience Methods, 250, 47-63.
Delorme, A., & Makeig, S. (2004). EEGLAB: An open source toolbox for analysis of
single-trial EEG dynamics including independent component analysis. Journal
of Neuroscience Methods, 134(1), 9-21.
AMBIGUITY DISADVANTAGE 46
De Cat, C., Klepousniotou, E., & Baayen, R. H. (2015). Representational deficit or
processing effect? An electrophysiological study of noun-noun compound
processing by very advanced L2 speakers of English. Frontiers in
Psychology, 6, 77.
De Rosario-Martínez, H. (2015). Phia: Post-hoc interaction analysis [R package].
Retrieved from http://cran.r-project.org/web/packages/phia.
Dholakia, A., Meade, G., & Coch, D. (2016). The N400 elicited by homonyms in
puns: Two primes are not better than one. Psychophysiology, 53(12), 1799-
1810.
Duffy, S. A., Morris, R. K., & Rayner, K. (1988). Lexical ambiguity and fixation times
in reading. Journal of Memory and Language, 27(4), 429-446.
Eddington, C. M., & Tokowicz, N. (2015). How meaning similarity influences
ambiguous word processing: The current state of the literature. Psychonomic
Bulletin & Review, 22(1), 13-37.
Federmeier, K. D., & Laszlo, S. (2009). Time for meaning: Electrophysiology
provides insights into the dynamics of representation and processing in
semantic memory. In B. H. Ross (Ed.), Psychology of learning and motivation
(pp. 1-44). San Diego, CA: Elsevier.
Frazier, L., & Rayner, K. (1990). Taking on semantic commitments: Processing
multiple meanings vs. multiple senses. Journal of Memory and
Language, 29(2), 181-200.
Frost, R., & Bentin, S. (1992). Processing phonological and semantic ambiguity:
Evidence from semantic priming at different SOAs. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 18(1), 58-68.
AMBIGUITY DISADVANTAGE 47
Grindrod, C. M., Garnett, E. O., Malyutina, S., & den Ouden, D. B. (2014). Effects of
representational distance between meanings on the neural correlates of
semantic ambiguity. Brain and Language, 139, 23-35.
Gottlob, L. R., Goldinger, S. D., Stone, G. O., & Van Orden, G. C. (1999). Reading
homographs: Orthographic, phonologic, and semantic dynamics. Journal of
Experimental Psychology: Human Perception and Performance, 25(2), 561-
574.
Hargreaves, I. S., Pexman, P. M., Pittman, D. J., & Goodyear, B. G. (2011).
Tolerating ambiguity: Ambiguous words recruit the left inferior frontal gyrus in
absence of a behavioral effect. Experimental Psychology, 58(1), 19-30.
Haro, J., Demestre, J., Boada, R., & Ferré, P. (2017). ERP and behavioral effects of
semantic ambiguity in a lexical decision task. Journal of Neurolinguistics, 44,
190-202.
Harpaz, Y., Lavidor, M., & Goldstein, A. (2013). Right semantic modulation of early
MEG components during ambiguity resolution. Neuroimage, 82, 107-114.
Hino, Y., Pexman, P. M., & Lupker, S. J. (2006). Ambiguity and relatedness effects in
semantic tasks: Are they due to semantic coding? Journal of Memory and
Language, 55(2), 247-273.
Hoffmann, S., & Evert, S. (2006). BNCweb (CQP-edition): The marriage of two
corpus tools. In S. Braun, K. Kohn, & J. Mukherjee (Eds.), Corpus Technology
and Language Pedagogy: New Resources, New Tools, New Methods (pp. 177-
195), Frankfurt: Peter Lang.
AMBIGUITY DISADVANTAGE 48
Hoffman, P., & Woollams, A. M. (2015). Opposing effects of semantic diversity in
lexical and semantic relatedness decisions. Journal of Experimental
Psychology: Human Perception and Performance, 41(2), 385-402.
Jager, B., & Cleland, A. A. (2015). Connecting the research fields of lexical ambiguity
and figures of speech: Polysemy effects for conventional metaphors and
metonyms. The Mental Lexicon, 10(1), 133-151.
Kawamoto, A. H. (1993). Nonlinear dynamics in the resolution of lexical ambiguity: A
parallel distributed processing account. Journal of Memory and Language,
32(4), 474-516.
Klepousniotou, E. (2002). The processing of lexical ambiguity: Homonymy and
polysemy in the mental lexicon. Brain and Language, 81(1), 205-223.
Klepousniotou, E., & Baum, S. R. (2007). Disambiguating the ambiguity advantage
effect in word recognition: An advantage for polysemous but not homonymous
words. Journal of Neurolinguistics, 20(1), 1-24.
Klepousniotou E., Gracco V.L., & Pike G.B. (2014). Pathways to lexical ambiguity:
fMRI evidence for bilateral fronto-parietal involvement in language processing.
Brain and Language, 131, 56-64.
Klepousniotou, E., Pike, G. B., Steinhauer, K., & Gracco, V. (2012). Not all
ambiguous words are created equal: An EEG investigation of homonymy and
polysemy. Brain and Language, 123(1), 11-21.
Klepousniotou, E., Titone, D., & Romero, C. (2008). Making sense of word senses:
The comprehension of polysemy depends on sense overlap. Journal of
Experimental Psychology: Learning, Memory, and Cognition, 34(6), 1534-1543.
AMBIGUITY DISADVANTAGE 49
Kornrumpf, B., Niefind, F., Sommer, W., & Dimigen, O. (2016). Neural correlates of
word recognition: A systematic comparison of natural reading and rapid serial
visual presentation. Journal of Cognitive Neuroscience, 28(9), 1374-1391.
Kutas, M., & Federmeier, K. D. (2011). Thirty years and counting: finding meaning in
the N400 component of the event-related brain potential (ERP). Annual Review
of Psychology, 62, 621-647.
Lee, C. L., & Federmeier, K. D. (2006). To mind the mind: An event-related potential
study of word class and semantic ambiguity. Brain Research, 1081(1), 191-202.
Lee, C. L., & Federmeier, K. D. (2009). Wave-ering: An ERP study of syntactic and
semantic context effects on ambiguity resolution for noun/verb
homographs. Journal of Memory and Language, 61(4), 538-555.
Leinenger, M., Myslín, M., Rayner, K., & Levy, R. (2017). Do resource constraints
affect lexical processing? Evidence from eye movements. Journal of Memory
and Language, 93, 82-103.
Loftus, G. R., & Masson, M. E. (1994). Using confidence intervals in within-subject
designs. Psychonomic Bulletin & Review, 1(4), 476-490.
Lucas, M. (2000). Semantic priming without association: A meta-analytic
review. Psychonomic Bulletin & Review, 7(4), 618-630.
MacGregor, L. J., Bouwsema, J., & Klepousniotou, E. (2015). Sustained meaning
activation for polysemous but not homonymous words: Evidence from
EEG. Neuropsychologia, 68,126-138.
MacGregor, L. J., Rodd, J. M., Gilbert, R. A., Hauk, O., Soboglu, E., & Davis, M. H.
(2020). The neural time course of semantic ambiguity resolution in speech
comprehension. Journal of Cognitive Neuroscience, 32(3), 403-425.
AMBIGUITY DISADVANTAGE 50
Maciejewski, G., & Klepousniotou, E. (2016). Relative meaning frequencies for 100
homonyms: British eDom norms. Journal of Open Psychology Data, 4, e6.
Maciejewski, G., Rodd, J. M., Mon-Williams, M., & Klepousniotou, E. (2019). The
cost of learning new meanings for familiar words. Language, Cognition and
Neuroscience, 1-23.
Matuschek, H., Kliegl, R., Vasishth, S., Baayen, H., & Bates, D. (2017). Balancing
Type I error and power in linear mixed models. Journal of Memory and
Language, 94, 305-315.
Mirman, D., Strauss, T. J., Dixon, J. A., & Magnuson, J. S. (2010). Effect of
representational distance between meanings on recognition of ambiguous
spoken words. Cognitive Science, 34(1), 161-173.
Mollo, G., Jefferies, E., Cornelissen, P., & Gennari, S. P. (2018). Context-dependent
lexical ambiguity resolution: MEG evidence for the time-course of activity in left
inferior frontal gyrus and posterior middle temporal gyrus. Brain and
Language, 177, 23-36.
Nelson, D. L., McEvoy, C. L., & Schreiber, T. A. (2004). The University of South
Florida free association, rhyme, and word fragment norms. Behavior Research
Methods, Instruments, & Computers, 36(3), 402-407.
Novick, J. M., Kan, I. P., Trueswell, J. C., & Thompson-Schill, S. L. (2009). A case
for conflict across multiple domains: Memory and language impairments
following damage to ventrolateral prefrontal cortex. Cognitive Neuropsychology,
26(6), 527-567.
AMBIGUITY DISADVANTAGE 51
Pexman, P. M., Hino, Y., & Lupker, S. J. (2004). Semantic ambiguity and the
process of generating meaning from print. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 30(6), 1252-1270.
Piercey, C. D., & Joordens, S. (2000). Turning an advantage into a disadvantage:
Ambiguity effects in lexical decision versus reading tasks. Memory &
Cognition, 28(4), 657-666.
Pollatsek, A., & Well, A. D. (1995). On the use of counterbalanced designs in
cognitive research: A suggestion for a better and more powerful
analysis. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 21(3), 785-794.
R Development Core Team (2004). R: A language and environment for statistical
computing. Vienna: R Foundation for Statistical Computing.
Rayner, K., & Duffy, S. A. (1986). Lexical complexity and fixation times in reading:
Effects of word frequency, verb complexity, and lexical ambiguity. Memory &
Cognition, 14(3), 191-201.
Rice, C. A., Beekhuizen, B., Dubrovsky, V., Stevenson, S., & Armstrong, B. C.
(2019). A comparison of homonym meaning frequency estimates derived from
movie and television subtitles, free association, and explicit ratings. Behavior
Research Methods, 51(3), 1399-1425.
Rodd, J. M. (2018). Lexical ambiguity. In S.-A. Rueschemeyer, & M. G. Gaskell
(Eds.) Oxford handbook of psycholinguistics (2nd Edition, pp. 120-144). Oxford,
UK: Oxford University Press.
AMBIGUITY DISADVANTAGE 52
Rodd, J. M., Gaskell, G., & Marslen-Wilson, W. (2002). Making sense of semantic
ambiguity: Semantic competition in lexical access. Journal of Memory and
Language, 46(2), 245-266.
Rodd, J. M., Gaskell, M. G., & Marslen-Wilson, W. D. (2004). Modelling the effects of
semantic ambiguity in word recognition. Cognitive Science, 28(1), 89-104.
Rodd, J. M., Johnsrude, I. S., & Davis, M. H. (2012). Dissociating frontotemporal
contributions to semantic ambiguity resolution in spoken sentences. Cerebral
Cortex, 22(8), 1761-1773.
Schneider, W., Eschman, A., and Zuccolotto, A. (2012). E-Prime 2.0. Pittsburgh, PA:
Learning Research and Development Center, University of Pittsburgh.
Seidenberg, M. S. (2007). Connectionist models of reading. In M. G. Gaskell (Ed.),
Oxford handbook of psycholinguistics (pp. 235-250). Oxford, UK: Oxford
University Press.
Sharbrough, F., Chatrian, G.-E., Lesser, R. P., Lüders, H., Nuwer, M., & Picton, T.
W. (1991). American Electroencephalographic Society guidelines for standard
electrode position nomenclature. Journal of Clinical Neurophysiology, 8, 200-
202.
Swaab, T., Brown, C., & Hagoort, P. (2003). Understanding words in sentence
contexts: The time course of ambiguity resolution. Brain and Language, 86(2),
326-343.
Simpson, G. B., & Burgess, C. (1985). Activation and selection processes in the
recognition of ambiguous words. Journal of Experimental Psychology: Human
Perception and Performance, 11(1), 28-39.
AMBIGUITY DISADVANTAGE 53
Simpson, G. B., & Krueger, M. A. (1991). Selective access of homograph meanings
in sentence context. Journal of Memory and Language, 30(6), 627-643.
Taler, V., Kousaie, S., & Lopez Zunini, R. (2013). ERP measures of semantic
richness: The case of multiple senses. Frontiers in Human Neuroscience, 7, 5.
Thompson-Schill, S. L., D’Esposito, M., Aguirre, G. K., & Farah, M. J. (1997). Role of
left inferior prefrontal cortex in retrieval of semantic knowledge: A re-
evaluation. Proceedings of the National Academy of Sciences, 94(26), 14792-
14797.
Twilley, L. C., & Dixon, P. (2000). Meaning resolution processes for words: A parallel
independent model. Psychonomic Bulletin & Review, 7(1), 49-82.
Twilley, L. C., Dixon, P., Taylor, D., & Clark, K. (1994). University of Alberta norms of
relative meaning frequency for 566 homographs. Memory & Cognition, 22(1),
111-126.
Vitello, S., & Rodd, J. M. (2015). Resolving semantic ambiguities in sentences:
Cognitive processes and brain mechanisms. Language and Linguistics
Compass, 9(10), 391-405.
Windisch Brown, S. (2008). Polysemy in the mental lexicon. Colorado Research in
Linguistics, 21(1), 1-12.
Witzel, J., & Forster, K. (2014). Lexical co-occurrence and ambiguity
resolution. Language, Cognition and Neuroscience, 29(2), 158-185.
AMBIGUITY DISADVANTAGE 54
Tables
Table 1. Examples of prime-target word pairs used in Experiments 1 and 2.
Prime Related target Unrelated target
HF-meaning LF-meaning A B Balanced homonym
fan-cheer fan-breeze fan-snake fan-cancel
Unbalanced homonym
pen-ink pen-farmer pen-yeast pen-add
Non-homonym1
fake-truth fake-fraud fake-expand fake-fetch
Non-homonym2
fur-fox fur-rabbit fur-chain fur-pill
AMBIGUITY DISADVANTAGE 55
Footnotes
1 RT analyses involving log transformation but not SD-based trimming produced
qualitatively similar results.
2 We report the results for Block and discuss why our experiments did not show
practice/prime-repetition effects in Section 2.1 in the Supplementary Material.
3 We began analysis with a model that included random intercepts and tested all
possible slopes for inclusion separately. Out of significant slopes, we first added the
most influential one (based on the value of ぬ2 from model-comparison tests) to the
base model and then tested whether the second most influential slope further
improved the model. We continued to test and include slopes until the model failed to
converge.
4 To determine the number of balanced and unbalanced homonyms in Pexman et
al.’s study (2004, Experiments 1-4), we used Twilley, Dixon, Taylor, and Clark’s
(1994) meaning-frequency ratings in Canadian English - the dialect spoken by the
recruited participants. We found that half of the homonyms had a highly dominant
meaning (i.e., meaning frequencies for these words differed by 41-79%), which
supports our claim that the study used both balanced and unbalanced homonyms
but did not distinguish between them. Note, however, that these are estimates only,
in that there may be slight differences in meaning-frequency ratings depending on
whether they are derived from television subtitles, word associations, or explicit
judgements (see Rice, Beekhuizen, Dubrovsky, Stevenson, & Armstrong, 2019).
AMBIGUITY DISADVANTAGE 56
5 We used BNCweb (CQP-edition; Hoffmann & Evert, 2006) to examine how often
primes and targets co-occurred within spoken and written language, up to four words
apart. This analysis confirmed that all but three related targets were rarely used
together with primes in natural discourse, and that our stimuli were not lexical
associates.
6 On average, only two of the 28 word pairs in each condition were forward- (e.g.,
“tent” in response to “camp”) or backward-generated associates (e.g., “camp” in
response to “tent”) in the University of South Florida Free Association Norms
(Nelson, McEvoy, & Schreiber, 2004). This indicates that primes and targets,
regardless of the condition, did not elicit each other’s meanings in a typical,
straightforward way.
7 Note that the scalp topography of ERPs does not allow us to make definitive claims
about the localization of neural sources.
8 The results for unbalanced homonyms were qualitatively similar after removing
some of the unbalanced homonyms with lesser-known LF meanings (see Section
2.4 in the Supplementary Material).
AMBIGUITY DISADVANTAGE 57
Figure Captions
Figure 1. Subject means of % error rates and untransformed RTs in ms for related
trials in Experiments 1a (Panel A) and 1b (Panel B). Error rates show 95%
confidence intervals adjusted to remove between-subjects variance (Loftus &
Masson, 1994).
Figure 2. Subject means of % error rates and untransformed RTs in ms for unrelated
trials in Experiments 1a (Panel A) and 1b (Panel B). Error rates show 95%
confidence intervals adjusted to remove between-subjects variance.
Figure 3. Subject means of % error rates and untransformed RTs in ms for related
(Panel A) and unrelated trials (Panel B) in Experiment 2. Error rates show 95%
confidence intervals adjusted to remove between-subjects variance.
Figure 4. Grand average waveforms for balanced/unbalanced homonyms and non-
homonyms during prime presentation (at major frontal, central, & posterior locations).
Negative amplitudes are plotted downwards.
Figure 5. Grand average waveforms for the HF-meaning, LF-meaning, and unrelated
targets of balanced homonyms during target presentation (at major frontal, central, &
posterior locations). Negative amplitudes are plotted downwards.
AMBIGUITY DISADVANTAGE 58
Figure 6. Grand average waveforms for the HF-meaning, LF-meaning, and unrelated
targets of unbalanced homonyms during target presentation (at major frontal, central,
& posterior locations). Negative amplitudes are plotted downwards.
AMBIGUITY DISADVANTAGE 59
Appendix
Sets of prime-target words pairs used in Experiments 1 and 2.
Prime Related pairs Unrelated pairs
HF target LF target Target A Target B
Balanced homonym
bay creek alcove tune ride bust breast burst basil eat
calf knee cattle trench bitter camp tent gay lag quick fan cheer breeze snake cancel forge advance hammer bird pig jam knife tight oval devil lean bend slim crime roar novel poem unique wipe reward palm wrist exotic sing mile pine oak desire cloak stroll plot writer acre curl plug prop pillar actor parrot dinner pupil lesson lens enter pan rank fifth odour device rift scrap pieces argue castle beach seal swim glue rapid monk shed hut skin fight dance squash sports potato alive anchor stall delay sell lip veil strip naked ribbon pond eagle tap sink knock beans poet temple chapel brow swan album tend habit nurse begin insect tense stress grammar cook tea toast dish beer skull ache utter aloud absolute fence sister yard grass inch invite betray
Unbalanced homonym
angle maths fisher bronze laugh cape jacket ocean error mental
AMBIGUITY DISADVANTAGE 60
chord song circle zoo sore corn crop toe preach quit ear listen cereal shelf excess egg goose urge boot ankle fleet navy swift smart ale flock herd fabric screen skill fry butter infant clay sign hide buried animal cheap acid host guest plenty sand throat lock shut comb pest saint mate pal chess galaxy crust mint ginger coin chin mess pad cloth foot anger frozen pen ink farmer yeast add pit dig cherry gaze sting pool bath resource tongue blade pulse vein seed milk gender pump flow shoes hunt jaw rail barrier protest willow foam ray shine fish ripe coal sheer thin veer fridge nose spray mist flower rival pigeon stern strict boat gift bin toll levy bell focus mud verse poetry tutor wet jungle wax warm moon dog heaven
Non-homonym Set 1
bald hairy wig vocal ton bulk huge vast wait funny
crew squad crowd arrow snow curve chart graph guard flood drain dry liquid banner prince fake truth fraud expand fetch fat broad tiny click witch fee wage permit mummy truce foster assist aid cash sick gap cavity hole whip ward grain wheat rice fairy exit grin teeth glad folder queen heap stack gather dwarf quote
AMBIGUITY DISADVANTAGE 61
hit shield slap reader prefer hook sharp trout busy neck hurdle bounce skip duke echo mask hat hood tide canoe raid rob troops vase clown saddle pony camel angel frown scan copy print beak shout elbow muscle bone envy loud shade shadow tree kiss mug silk linen shiny cheese rage slice divide sword ghost active smash crush grind worm virus tall giant height code worry trim barber beard bag spoon wool yarn goat bread foe
Non-homonym Set 2
abuse harm cruel menu chalk bet luck gamble parent collar
burn grill heat hint famous dawn dusk bright rebel toss deaf blind noise purse golf dip plunge rinse dragon humble drift wander yacht comedy gun feast supper cake smooth horn fog cloud rain scream hug fur fox rabbit chain pill grasp grab snatch melt trial hay farm nest pearl resist honey sauce sweet fun rugby leap runner jump owl powder load cargo lorry tour rub loop rope shape sniff tribe peak hill climb batch bug pilot sky cabin dirt tape push hurt ram rat snack ritual pray cult stew honest rod copper cane era pillow smoke vapour oven dairy twin sour apple candy bullet weapon spy agent enemy pale toad
AMBIGUITY DISADVANTAGE 62
teach guide learn escape edge tin bottle metal sad track torch cave lamp speed scalp void null valid island rural